Every line of code writes a history of power. So does every foundry yield table.
On July 25, 2026, Samsung Electronics and Broadcom signed a memorandum of understanding at the AI Summit in San Francisco. The public figure attached to that memorandum exceeds $200 billion in combined memory and foundry services through 2030. The public narrative is about vertical integration: HBM4, HBM4E, 2nm logic, and advanced packaging in a single supply offer. TSMC, a pure-play foundry, cannot copy that bundle because it does not make memory. That is structurally true. A non-binding MOU, however, is not a manufacturing result. The distinction matters. I have spent years auditing governance frameworks where the whitepaper promised decentralization and the code delivered admin keys. The same discipline applies here. The deal is not the question. The yield is.
This agreement is not a normal procurement contract. It is an attempt to redefine the geometry of the AI semiconductor supply chain. Samsung is betting that the complexity of modern AI silicon will favor suppliers who control the full stack over specialists who only touch logic. TSMC can manufacture the chip. Samsung is offering to manufacture the chip, supply the memory, and handle the packaging integration—three operations that currently require three different vendors. That is the structural argument. It is also a direct challenge to the assumption that one Taiwanese company must remain the unaccountable choke point of AI compute. The broader market is moving toward custom silicon. ASIC shipments are growing 44.6 percent year over year, more than double the 16.1 percent growth of merchant GPUs. Custom ASICs now represent 27.8 percent of AI server shipments. Broadcom, with roughly 60 percent of the custom AI ASIC co-design market and a $73 billion AI backlog, is the demand vector that can turn Samsung's thesis into a test.
Let me be clear about what is being tested. Samsung's foundry market share sits at around 7 to 8 percent. TSMC commands 72 to 73 percent of global foundry revenue and roughly 95 percent of the AI accelerator space. Over the past three years, Samsung has tried to close that gap by marketing gate-all-around transistors and aggressive process roadmaps. The marketing succeeded. The silicon did not. Samsung's 2nm process currently yields between 50 and 60 percent, while TSMC maintains yields of 80 percent or higher. For a custom AI ASIC, yield is not a footnote. It is the price. A 20-to-30-percentage-point yield gap at the leading node translates directly into higher per-wafer costs, lower throughput, and longer qualification cycles. No amount of vertical integration can erase that arithmetic.
We didn't need another MOU to confirm that AI demand is real. The power draw of the hyperscalers told us that two years ago. What we needed was evidence that Samsung can execute at the frontier of logic manufacturing. So far, that evidence has not appeared in any audit I have read. In my work with DeFi protocol governance, I have learned to compare a proposal's promise with the contract's actual access controls. The equivalent exercise here is to compare Samsung's foundry roadmap with its shipped wafers. The roadmap is ambitious. The wafers are not yet competitive.
The Q2 2026 earnings tell the same story from a different direction. Samsung's memory business delivered ₩120.8 trillion in revenue, a 471 percent year-over-year increase driven by the HBM supercycle. HBM4 sales are projected to triple in the third quarter, and HBM4E samples are already in the hands of major customers. These are extraordinary numbers. And after the earnings call, Samsung shares fell 13.4 percent. Markets rarely punish strength in memory alone. They punished the gap between the memory story and the foundry story. Investors can see that the Broadcom MOU is not a revenue contract; it is a subsidy plan. The foundry push is being funded by memory profits, and the yield table determines how long that subsidy must continue.
Broadcom's own arithmetic explains why the MOU is structured the way it is. Broadcom is managing a $73 billion AI backlog and has set a target of $100 billion in annualized AI revenue by fiscal 2027. It cannot afford to remain dependent on a single manufacturer for custom AI silicon. Charlie Kawwas, Broadcom's president of semiconductor solutions, said: "As AI infrastructure continues to scale, close collaboration across the semiconductor ecosystem becomes increasingly important." That is the public rationale. The private rationale is simpler. Broadcom wants pricing leverage, capacity optionality, and a credible second source for logic and packaging. Samsung wants the brand of becoming Broadcom's integrated foundry-memory partner. The MOU serves both interests, which is precisely why it should be classified as a strategic hedge rather than a conversion.
To understand why the yield gap is decisive, look at Broadcom's customer list. The hyperscalers in that portfolio — Google, Meta, OpenAI — do not buy wafers for fun. They buy silicon to serve inference traffic that must run at a predictable cost. A custom AI ASIC is designed for a specific workload, and its lifetime value is computed from throughput per watt and cost per chip. If Samsung's 2nm process yields 55 percent, Broadcom must manufacture roughly two wafers to get the same number of good dies that TSMC would produce from one and a half wafers. The extra wafers consume manpower, cleanroom time, and packaging capacity. The yield delta is not an abstract engineering metric; it becomes a line item in the hyperscaler's operating expenses. That is why Broadcom said "close collaboration" instead of "we have made a final choice." The collaboration is real. The commitment is conditional.
Consider the inside of Samsung's offer. HBM4 and HBM4E are the memory components. The 2nm process node is the logic component. The 2.3D and 2.5D advanced packaging is the integration layer. Bundling these into one agreement seems like a superior governance structure because it collapses three counterparties into one. But there is a hidden cost: the same counterparty also controls the failure modes. If Samsung's 2nm yield lags, Broadcom cannot simply shift the logic to TSMC while retaining Samsung memory, because the packaging integration crosses the same division that is struggling to produce logic. That is not a single point of failure; it is a triangle of failure sharing one balance sheet.
This is where the conventional "TSMC cannot match the bundle" narrative misses the real control structure. TSMC does not need to match the bundle. TSMC needs to continue beating Samsung on logic yield, and then use its own advanced packaging capacity to integrate whatever memory Broadcom chooses to buy elsewhere. The pure-play model is weaker in theory and stronger in practice, because it does not allow a memory division's exceptional margins to mask a foundry division's poor execution. Governance isn't a boardroom abstraction. It is the set of constraints that keeps one unit from being funded indefinitely by another unit without producing competitive output.
We didn't reach that conclusion by trusting press releases. I have been through enough audits to know that most structural arguments fail at the boundary between divisions. In DeFi, we call it composability risk. In a conglomerate, it is called transfer pricing. Samsung's memory and foundry units are in the same building, but they do not share the same cost curve. HBM profits are real because Samsung controls the memory supply chain from design to production. Foundry losses are real because Samsung has not matched TSMC's leading edge. The Broadcom MOU does not change that cost curve. It only gives Samsung more time to fix it, paid for by memory revenue.
Young Hyun Jun, Vice Chairman and CEO of Samsung's DS Division, framed the bet in appropriately grand terms: "AI is driving unprecedented demand for tightly integrated semiconductor technologies spanning memory, logic and advanced packaging. By combining Samsung's memory and foundry expertise with Broadcom's AI and connectivity leadership, we aim to continue to deliver technologies that power the next generation of AI infrastructure." The rhetoric is coherent. The yield data is not yet coherent. And the MOU is non-binding, which means the $200 billion figure represents a maximum addressable market if Samsung meets its roadmap, not a contractual commitment from Broadcom. Actual volume will depend on future product roadmaps, qualification results, and the willingness of hyperscalers to accept a second source with a meaningful yield gap.
The contrarian reading is sharp. Vertical integration may be the wrong tool for the problem Samsung faces. The problem is not lack of customer interest. The problem is lack of process maturity. Samsung could solve that with more engineering resources, more time, and more transparency. Instead, the company is trying to solve it with a commercial agreement. The agreement secures a potential customer, but it does not move the needle on the physical process line. To put it in language I use when reviewing smart contracts: the MOU is an authorization, not an execution. It grants the right to run the transaction. It does not guarantee the transaction succeeds. Every line of code writes a history of power, but the power is worthless if the underlying state transition fails.
There is also a deeper question about how AI infrastructure should be governed. The demand for AI compute is increasingly concentrated in a handful of hyperscalers. The supply of leading edge logic is concentrated in a handful of foundries. A non-binding MOU between Samsung and Broadcom is not a decentralization event. It is a reconfiguration of centralization. Broadcom gains flexibility. Samsung gains scale. The hyperscalers may gain a pricing alternative. But the systemic risk remains: the world's most important economic infrastructure will still depend on a small number of factories, a small number of water sources, and a small number of executive decisions. For the crypto market, this is not a metaphor. Every AI agent executing on-chain draws power from this exact supply chain, and if the chain centralizes, the agent economy inherits its failure modes. That is a governance problem, not just a supply chain problem. We should hold Samsung to the same standard we hold a protocol. Does it publish honest yield data? Does it identify the failure modes before the market does? Does it treat the foundry as a public utility rather than a strategic experiment?
The answer today is no. Samsung does not publish granular 2nm yield data. It publishes aspirational statements about technology leadership. In a market where a MOU can attract $200 billion in headline attention, opacity is a competitive advantage. But it is a destructive one. TSMC's lead is not only a matter of process technology. It is a matter of trust built from years of verifiable execution. Every line of code writes a history of power, and every whispered chip rumor writes a history of fear. If Samsung wants to convert this MOU into actual foundry share, it needs to start publishing yield numbers with the same regularity it publishes memory earnings. That is the real test.
The next step is not a press conference. It is the Pyeongtaek campus, where Samsung's 2nm line must prove that the yield curve can move from the mid-50s to the high-70s within the next four quarters. If it does, the Broadcom MOU becomes a credible foundation for a new supply chain architecture. If it does not, the $200 billion figure will be remembered as the largest non-binding wish in semiconductor history. Broadcom knows this. The market knows this. Samsung's leadership knows this. The knowledge is not the bottleneck. Execution is the bottleneck.
Truth emerges from transparency, not from silence. For too long, the semiconductor industry has treated yield data as a competitive secret rather than a governance requirement. That era is ending. AI infrastructure is too important to be governed by rumor. Investors should demand quarterly yield disclosures the way they demand audit reports. Regulators should ask why the most critical fabrication capacity in the world is still treated as a black box. And Samsung should understand that its next report is the one nobody wants to write. I will be watching for a single sentence: the honest number on 2nm yield. That sentence will determine whether this is the beginning of a TSMC challenge or the most expensive hedge in semiconductor history.